A pipeline early leakage detection system based on multi-sensor network

Abstract

The application relates to the technical field of pipeline detection, and discloses a pipeline early leakage detection system based on a multi-sensor network, which comprises a multi-sensor network for acquiring multi-dimensional sensing signals; an edge gateway for receiving the signals and performing preprocessing to generate standardized preprocessing signals corresponding to each physical dimension; a first processor for extracting features from the standardized preprocessing signals respectively and constructing corresponding feature vectors; a second processor for performing space-time alignment and splicing on the feature vectors, generating comprehensive fusion feature vectors and transmitting the comprehensive fusion feature vectors to an analysis platform; and the analysis platform for inputting the comprehensive fusion feature vectors into a hybrid deep neural network for inference determination and outputting a leakage state label of a pipeline to be detected; and the analysis platform performing alarm and dispatching strategies according to the leakage state label. Through multi-dimensional physical parameter fusion and multi-stage determination and verification, the false alarm rate caused by interference is reduced, and the accuracy and stability of pipeline leakage state discrimination are improved.

Description

Technical Field

[0001] This invention relates to the field of pipeline inspection technology, specifically to a pipeline early leakage detection system based on a multi-sensor network. Background Technology

[0002] As an infrastructure for transporting fluid media, pipelines are susceptible to leakage due to factors such as material aging, environmental corrosion, and geological subsidence during long-term operation. This can lead to the formation of micropores or cracks in the pipe walls. Existing pipeline monitoring technologies typically involve deploying sensors along the pipeline to collect operational data and determine its health status.

[0003] Traditional pipeline inspection methods often rely on single-dimensional physical parameters, such as using only the characteristics of pressure fluctuations within the pipe or a single acoustic emission signal for threshold comparison. However, actual pipeline operation is often accompanied by pump start-up and shutdown vibrations, fluid turbulence, and external environmental background noise. These fluctuations under normal operating conditions can easily mask early, subtle leak signals. Single-signal-based judgment methods have limitations in setting alarm thresholds. If the threshold is set too low, interference from normal operating conditions can trigger false alarms; if the threshold is set too high, it becomes difficult to detect early, weak leak characteristics.

[0004] With the development of monitoring equipment, some existing monitoring systems have begun to incorporate multiple types of sensors. However, data processing often remains at the level of independent analysis of each parameter, lacking strict alignment and feature fusion of multi-source heterogeneous data such as pressure, vibration, and temperature in both time and space dimensions, and failing to utilize the coupling relationships between different physical phenomena. Furthermore, the judgment models of existing systems typically rely on single sampling results for output. When faced with transient interference signals or missing sensing data due to communication issues, the judgment results are prone to abrupt changes. This judgment architecture, lacking fault-tolerant verification mechanisms and sliding tolerance processing, limits the stability of the monitoring system and makes it difficult to meet the engineering requirements for accurate state determination in industrial settings. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a pipeline early leakage detection system based on a multi-sensor network. This system solves the technical problem that existing pipeline monitoring methods rely on a single physical parameter and lack joint analysis and fault-tolerant verification mechanisms for multi-source heterogeneous data, resulting in limited accuracy in identifying early and minor leakage under complex operating conditions and a tendency to generate false alarms.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a pipeline early leakage detection system based on a multi-sensor network. The system includes: a multi-sensor network deployed on the pipeline, comprising multiple monitoring nodes along the pipeline to be detected, each monitoring node having a pressure sensor, a vibration sensor, an acoustic emission sensor, and a temperature sensor fixedly installed therein to acquire corresponding pressure signals, vibration signals, acoustic emission signals, and temperature signals; an edge gateway electrically connected to the multi-sensor network, used to receive the pressure signals, vibration signals, acoustic emission signals, and temperature signals and perform preprocessing to generate standardized preprocessed signals corresponding to each physical dimension; and a first processor, connected to... The edge gateway connection is used to extract features from the standardized preprocessed signal based on a sliding time window, and construct pressure feature vector, vibration feature vector, acoustic emission feature vector, and temperature feature vector. The second processor, electrically connected to the first processor, is used to perform spatiotemporal alignment and splicing of each feature vector to generate a comprehensive fusion feature vector and transmit it to the analysis platform. The analysis platform is used to input the comprehensive fusion feature vector into a hybrid deep neural network for inference and judgment, and output a leakage status label for the pipeline to be detected. When the leakage status label indicates that there is a leakage, the analysis platform triggers a leakage location program and executes alarm and scheduling strategies based on the label.

[0008] Preferably, the edge gateway includes an adaptive filter, and the edge gateway performs preprocessing, including: extracting the time-domain signal sequence of each signal and performing discrete wavelet decomposition to obtain low-frequency approximation coefficient components and high-frequency detail coefficient components; using an adaptive soft thresholding algorithm to perform threshold quantization processing on the high-frequency detail coefficient components, and reconstructing a one-dimensional preliminary denoised signal in the time domain using the retained low-frequency approximation coefficient components and the quantized high-frequency detail coefficient components; using the background signal collected by the environmental reference sensor deployed in the non-monitoring area as the input reference noise signal vector, and inputting it together with the preliminary denoised signal into the minimum mean square adaptive filter; dynamically adjusting the tap weight coefficients through iterative optimization to approximate the statistical characteristics of the residual noise, and separating the residual noise from the preliminary denoised signal to output the physical signal after composite denoising processing; and performing maximum and minimum value normalization processing based on the physical signal to generate a standardized preprocessed signal.

[0009] Preferably, the first processor is specifically used for: calculating mean parameters, variance parameters, peak-to-peak values, and kurtosis parameters from the standardized preprocessed signal of the corresponding pressure signal and concatenating them to construct a pressure feature vector; performing a fast Fourier transform on the standardized preprocessed signal of the vibration signal to extract the centroid frequency, calculating the wavelet packet energy entropy using wavelet packet transform, and concatenating them to construct a vibration feature vector; setting a dynamic trigger threshold calculated based on the mean of the signal envelope curve for the standardized preprocessed signal of the corresponding acoustic emission signal to identify transient acoustic emission events, extracting impact count parameters, average event energy parameters, and average peak amplitude parameters, concatenating them according to a set arrangement order, and constructing a one-dimensional acoustic emission feature vector; and calculating the base temperature values ​​of each distributed temperature sensor node using the standardized preprocessed signal of the temperature signal, extracting the spatial maximum temperature difference parameter, spatial temperature gradient parameter, and spatial coefficient of variation parameter, and concatenating them to construct a temperature feature vector.

[0010] Preferably, the identification of the acoustic emission transient event includes: performing envelope detection on the standardized preprocessed signal of the acoustic emission signal using Hilbert transform to obtain the envelope curve; multiplying the arithmetic mean of the values ​​of all sampling points of the envelope curve within the current data processing window with a set trigger coefficient to calculate the dynamic trigger threshold; when the value of the envelope curve crosses the dynamic trigger threshold from small to large, marking the sampling number corresponding to the crossing point as the event start point; when the value of the envelope curve falls back from large to small and crosses the dynamic trigger threshold, and does not rise again within the set event locking time window, marking the sampling number corresponding to the falling back point as the event end point; the envelope curve data segment between the event start point and the event end point constitutes an independent acoustic emission transient event.

[0011] Preferably, the second processor is specifically used for: obtaining a global timestamp label from the edge gateway, the global timestamp label being generated by the edge gateway when reading each feature vector and generating the standardized preprocessing signal; aligning the time dimension using the system clock cycle of the current data processing window as a reference criterion; obtaining the spatial index incrementing order from the edge gateway, the spatial index incrementing order being generated by the edge gateway according to the distribution of the acquisition nodes along the axis of the pipeline to be detected; clustering the pressure, vibration, and acoustic emission feature vectors belonging to the same monitoring section coordinate into a local physical parameter set, adding the temperature feature vector reflecting the heat conduction state of the entire pipeline as a global physical parameter to the back region of the local physical parameter set, and generating a cross-dimensional comprehensive fusion feature vector by performing a cascade mapping operation through the data bus; and sending the comprehensive fusion feature vector to the analysis platform, the analysis platform being further used for: reading the global moving mean and global moving variance from the model parameter weight file for standardization, and introducing scaling and translation parameters to perform linear reconstruction.

[0012] Preferably, the hybrid deep neural network includes a one-dimensional convolutional neural network, a long short-term memory network, and a fully connected layer. The analysis platform is specifically used for: performing sliding matrix multiplication and addition operations on the batch-normalized feature vectors using the one-dimensional convolutional neural network with convolutional kernels of a set length to generate local feature mapping values; performing max pooling dimensionality reduction on the output local feature mapping matrix to generate a feature sequence; receiving the feature sequence through the long short-term memory network, resolving spatial structure dependencies through an internal control mechanism, obtaining the hidden layer output vector of the last sequence step, flattening it, and passing it to the fully connected layer; mapping the high-dimensional abstract features to classification nodes equal to the set number of leakage categories through the fully connected layer; calling a normalization exponential function to probabilistically map the output values ​​of the classification nodes, calculating the predicted probability of the pipeline under test being in each leakage state, and obtaining a probability distribution set; extracting the state label corresponding to the maximum predicted probability in the probability distribution set as the leakage state label of the pipeline under test.

[0013] Preferably, before extracting the status label corresponding to the maximum predicted probability, the analysis platform is further configured to: compare the maximum predicted probability value with a set confidence threshold; if it is greater than or equal to the threshold, output the corresponding category index as the leakage status label of the pipeline to be detected; if it is less than the threshold, mark the output result as unresolved, initiate a moving average determination mechanism, continuously acquire the probability distribution sets of the next two adjacent data processing windows, forming a moving determination sequence with a total length of three windows; calculate the arithmetic mean of the predicted probabilities corresponding to the same category to generate a smoothed probability distribution set, and compare the maximum smoothed predicted probability value with a set secondary determination threshold. If it is greater than or equal to the secondary determination threshold, output the corresponding status label; if it is still less than the secondary determination threshold, force the leakage status label of the pipeline to be detected to be output as normal operation status.

[0014] Preferably, the analysis platform triggers a leak location procedure, including: retrieving upstream and downstream pressure feature vectors and upstream and downstream acoustic emission feature vectors synchronously collected by edge gateways located upstream and downstream of the target pipe segment to be detected; performing time alignment analysis on the upstream and downstream pressure feature vectors to obtain the time difference, and calculating the time difference between the leak point and the upstream acquisition node based on the propagation speed of the pressure wave in the pipe fluid medium to locate the physical distance; extracting the effective values ​​of the acoustic emission signal within a set time window based on the upstream and downstream acoustic emission feature vectors as the signal amplitudes of the upstream and downstream nodes, using the physical mechanism that the energy of the acoustic emission wavefront attenuates with the propagation distance, constructing and solving an algebraic equation with the signal amplitudes of the upstream and downstream nodes and the acoustic emission attenuation coefficient as parameters, and calculating the attenuation location physical distance based on the attenuation model; assigning time difference location weights and attenuation location weights to the time difference location physical distance and the attenuation location physical distance respectively, performing linear multiplication and addition operations to generate a comprehensive weighted location distance, which is used as the final leak point coordinate output value.

[0015] Preferably, the analysis platform is connected to the host computer control center. The analysis platform executes alarm and scheduling strategies based on the leakage status label, including: when the leakage status label corresponds to a micropore leakage status, generating a level one early warning message and sending it to the host computer control center; when the leakage status label corresponds to a medium crack leakage status, sending an audible and visual alarm drive command to the host computer control center, and simultaneously sending commands to the upstream and downstream edge gateways to set the data sampling frequency to twice the original parameter; when the leakage status label corresponds to a severe pipe wall rupture status, triggering a three-level emergency interlocking mechanism, outputting a DC hard-wired interlocking signal to drive the pneumatic actuator or electric valve to perform a physical shut-off operation for pipeline isolation, starting a response confirmation timer and continuously monitoring the valve full-close limit feedback signal.

[0016] Preferably, the analysis platform is further configured to: when the physical parameter acquisition value within a local time window is lost before performing the batch normalization process, the analysis platform extracts historical feature elements of the same physical parameter acquired at the previous sampling time and the previous two sampling times; establish the differential equality relationship between three adjacent discrete sampling points, derive the prediction equation for the current time value, and obtain the repaired feature elements at the current time through first-order linear extrapolation calculation; fill the corresponding index positions of the original data matrix with the repaired feature elements to reconstruct the feature vector sequence; and substitute the constructed feature vector sequence into the comprehensive fusion feature vector to continue the subsequent processing flow.

[0017] This invention provides a pipeline early leakage detection system based on a multi-sensor network, which has the following advantages:

[0018] 1. Physical signals from four dimensions—pressure, vibration, acoustic emission, and temperature—are acquired through an edge gateway, spatiotemporally aligned and stitched together. A hybrid deep neural network, combining a one-dimensional convolutional neural network and a long short-term memory network, is then used to extract local spatial features and temporal dependencies from the data. Compared to single-parameter detection, this multi-parameter fusion method can comprehensively reflect pipeline state changes from different physical dimensions, improving the accuracy of early leakage detection.

[0019] 2. This invention introduces a confidence threshold verification and moving average determination mechanism after the deep neural network outputs the prediction result. When the maximum predicted probability is lower than the confidence threshold, the output is marked as undecided, and a smooth calculation is performed based on the probability distribution set of subsequent adjacent data processing windows before comparison. This mechanism reduces misjudgments caused by transient noise or background interference in a single sampling, lowers the false alarm rate, and improves the stability of the determination result.

[0020] 3. This invention designs a data loss repair operation and a hierarchical scheduling strategy. When local data acquisition is lost, a differential equality relationship is established using historical feature elements to perform first-order linear extrapolation calculation to fill in the missing features, ensuring the data continuity of subsequent model inference and the fault tolerance of the system; at the same time, different levels of alarm and scheduling actions are executed according to the specific leakage status determined, realizing closed-loop control from status perception to physical intervention. Attached Figure Description

[0021] Figure 1 A schematic diagram of the structure of a pipeline early leakage detection system based on a multi-sensor network provided in an embodiment of the present invention;

[0022] Figure 2 This is a comparison chart showing the identification accuracy of different algorithms of the present invention under various leakage conditions.

[0023] Figure 3 This is a graph showing the convergence performance analysis of the cross-entropy loss function in the closed-loop training process of this invention.

[0024] Figure 4 This is a verification diagram showing the improvement in leakage location accuracy achieved by the multi-source heterogeneous information collaborative weighting method of the present invention. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see the appendix Figure 1This invention provides a pipeline early leakage detection system based on a multi-sensor network. The system includes a multi-sensor network, an edge gateway, an analysis platform, and an alarm terminal. The multi-sensor network is deployed along the pipeline to be detected. The multi-sensor network is communicatively connected to the edge gateway, which is connected to a first processor. The first processor is connected to a second processor via a network interface and / or a wireless network. The second processor is communicatively connected to the analysis platform, and the analysis platform is communicatively connected to the alarm terminal. Specifically:

[0027] The multi-sensor network includes multiple monitoring nodes deployed along the pipeline to be inspected. Each monitoring node is equipped with a pressure sensor, a vibration sensor, an acoustic emission sensor, and a temperature sensor, which are used to acquire the corresponding pressure signal, vibration signal, acoustic emission signal, and temperature signal.

[0028] The multi-sensor network includes multiple monitoring nodes deployed along the pipeline to be inspected. The spacing between adjacent monitoring nodes is set to 10 to 20 meters. Each monitoring node is equipped with a pressure sensor, a vibration sensor, an acoustic emission sensor, and a temperature sensor. The multi-sensor network collects the electrical signals output by the four types of sensors in each monitoring node in real time based on a synchronous sampling clock, thereby acquiring pressure signals, vibration signals, acoustic emission signals, and temperature signals.

[0029] The edge gateway sets cutoff frequencies for the frequency response ranges of different types of sensors and performs filtering operations on each signal. After the filtering operation is completed, the edge gateway combines wavelet transform algorithm and least mean square adaptive filtering algorithm to perform denoising calculation on the signal. The edge gateway uses the maximum-minimum algorithm to uniformly map the denoised data of different dimensions to a set numerical range, generating a standardized preprocessed signal.

[0030] A first processor, connected to the edge gateway, for example, is used in a conventional intelligent monitoring device with data processing capabilities. This processor calculates and extracts time-domain statistical parameters from the pressure signal as pressure features based on a set sliding time window. It performs a Fast Fourier Transform on the vibration signal to extract frequency-domain energy distribution parameters as vibration features. For the acoustic emission signal, it sets a trigger threshold to extract effective transient waveform parameters as acoustic emission features. Based on the spatial distribution of the temperature sensor array, it calculates temperature difference values ​​to extract spatial distribution parameters as temperature features.

[0031] The second processor is electrically connected to the first processor. For example, the second processor is applied to another conventional intelligent monitoring device with data processing function. This device is wirelessly connected to the device where the first processor is located. It is used to read the feature data output by all monitoring nodes. According to the set sensor category order and node spatial arrangement order, the pressure feature, vibration feature, acoustic emission feature and temperature feature are sequentially spliced ​​to generate a multi-dimensional fusion feature vector. After splicing, the analysis platform performs batch normalization operation on the multi-dimensional fusion feature vector to adjust the mean and variance of the feature data components.

[0032] The analysis platform incorporates a multi-layered deep neural network model. The platform inputs fused feature vectors into this model, which then sequentially extracts local spatial features through a one-dimensional convolutional layer, extracts time-series features through a long short-term memory network layer, and finally outputs the probability values ​​of the pipeline's different leakage states to determine the current leakage level.

[0033] When the analysis platform determines that a leakage exists, it performs leakage point location estimation. The platform extracts the time difference data of the leakage physical signal reaching different monitoring nodes along the route, and combines this with the wave velocity parameters within the medium to calculate and establish a set of spatial distance equations based on the time difference. The platform then inverses the propagation distance based on the attenuation degree of the signal amplitude at different nodes. Finally, the platform performs a weighted fusion calculation of the time difference-calculated distance and the amplitude attenuation-calculated distance to obtain the absolute coordinates of the leakage point.

[0034] The alarm terminal receives leakage level data and location coordinate data output by the analysis platform, determines the alarm level corresponding to the current leakage level according to the preset response rules, triggers the corresponding audible and visual alarm device hardware command according to the alarm level, and renders and displays the leakage location coordinates and sensor status waveform data on the monitoring screen interface.

[0035] Multidimensional data acquisition is achieved using sensor modules deployed along the pipeline to be inspected. To acquire spatiotemporally correlated pipeline physical field operation data, the sensor modules employ a hierarchical and customized spatial deployment strategy. When determining the basic spatial topology of the monitoring nodes, conventional straight pipe sections are identified along the axis of the pipeline to be inspected. On these conventional straight pipe sections, the coordinates of the basic monitoring nodes are sequentially marked at intervals of 10 to 20 meters. This distance setting is used to strike a balance between covering monitoring blind spots and controlling hardware costs, ensuring that even minor leaks causing changes in the physical field can be captured by adjacent nodes.

[0036] Furthermore, the system identifies areas with complex pipeline structures and implements localized densification of monitoring nodes. Bends, reducers, valve flange connections, and tee junctions along the pipeline route are designated as key monitoring areas prone to leakage. Within these key monitoring areas, the density of monitoring nodes is increased, reducing the physical spacing between nodes to below a preset value, set at 3 to 5 meters. Increasing the node density improves the spatial resolution of data acquisition in areas with abrupt changes in local fluid dynamics conditions.

[0037] Physical installation and acoustic coupling matching were performed for the heterogeneous sensors inside the monitoring node. Each monitoring node corresponds to a spatial cross-sectional area along the pipeline. Sensors with different sensing principles were arranged within this area. Pressure sensors were fixedly installed on the pipeline sidewall to sense fluctuations in internal fluid pressure. Triaxial accelerometers and broadband acoustic emission sensors were fixed to the polished outer surface of the pipeline. When leakage occurs in the pipeline, the outward spray of water generates broadband vibrations, and the microscopic deformation of the pipe material releases transient elastic waves; these sensors are used to capture these mechanical wave signals. During installation, a rigid acoustic coupling agent was filled between the sensor base and the pipe wall surface to eliminate air gaps at the contact interface. This rigid connection suppresses signal attenuation and enables the transmission of broadband micro-vibrations and high-frequency transient elastic waves. Temperature sensors, using a thermocouple array structure, were linearly distributed along the axial extension and radial depth directions of the pipeline and buried in the surrounding soil layer to obtain the temperature gradient distribution within the three-dimensional space around the pipeline. Fluid leakage causes changes in the specific heat capacity and local thermal field of the surrounding soil medium. Temperature sensors help characterize the leakage state by capturing these temperature gradient changes.

[0038] Finally, engineered physical protection and insulation treatment are implemented for the monitoring nodes. To address the issues of water accumulation and dust interference in underground or high-damping enclosed monitoring environments, sealing and protection operations are performed on the sensor housings, data acquisition terminals, and communication cable interfaces adhering to the pipe walls. The protection operation employs a waterproof and dustproof encapsulation process to block external environmental interference with the transmission of weak electrical signals. Regarding the specific molding method of the sensor protective housing and the selection of waterproof and insulating potting compound materials, those skilled in the art can make conventional selections and configurations based on the temperature and humidity conditions at the pipeline site. The encapsulation process is well-known technology in this field and will not be elaborated upon here.

[0039] Multidimensional data acquisition is performed using a sensor array configured within the monitoring node to synchronously sense multi-physics changes caused by pipeline leakage. Pressure signal acquisition utilizes a microelectromechanical system (MEMS) pressure sensor. This sensor is fixed to the pipe sidewall, with a range of 0–1 MPa and a sampling frequency greater than or equal to 200 Hz. This sampling frequency is chosen because the pressure wave generated at the moment of pipe rupture is a high-frequency transient signal, and the system needs to satisfy the Nyquist sampling theorem to obtain a complete pressure waveform profile. At the physical sensing level, localized leakage in the pipe leads to fluid loss, disrupting the original fluid dynamics balance and causing pressure drop fluctuations in the local area. The MEMS pressure sensor contains a silicon piezoresistive pressure-sensitive diaphragm. The fluid pressure wave actuates the diaphragm, causing mechanical deformation and a change in the resistance of the attached pressure-sensitive resistor. A Wheatstone bridge circuit converts this resistance change into a voltage signal output, acquiring the static water pressure and dynamic pressure state characteristics inside the pipe.

[0040] Vibration signals are acquired using a triaxial accelerometer fixed to the pipe surface structure. The frequency response range of the triaxial accelerometer is set to 0–5 kHz. This frequency band is designed to cover the main frequency range of pipe wall resonance caused by water jetting. When leakage occurs in the pipe, the water jetting from the pores continuously scours the outer side of the pipe wall and the surrounding soil medium. The fluid impact force and the reaction force of the surrounding medium excite mechanical vibration on the pipe wall. The triaxial accelerometer is equipped with a mass block and an elastic sensing element. By measuring the force deformation of the mass block under inertia, it simultaneously acquires vibration acceleration components along three mutually perpendicular dimensions: axial, radial, and tangential. By acquiring vibration data in three spatial dimensions, the physical impact state caused by water jetting is reflected.

[0041] Acoustic emission signals are acquired using an acoustic emission sensor. The sensor's detection frequency band is set between 20kHz and 1MHz. This high-frequency range is chosen to avoid interference from low-frequency noise from normal mechanical operation and vehicle traffic, and is specifically designed to extract high-frequency signals generated by microscopic deformation. In the early stages of pipe leakage, high-pressure fluid passes through microscopic pores or cracks in the pipe wall, generating jet friction. Simultaneously, the pipe material undergoes microscopic deformation due to internal stress changes. These microscopic physical processes release transient elastic waves. The acoustic emission sensor contains a piezoelectric ceramic sensing element. When the transient elastic wave propagates along the pipe wall to the sensor base, the piezoelectric ceramic sensing element, based on the positive piezoelectric effect, converts the microscopic stress induced by the elastic wave into an alternating charge signal, capturing high-frequency acoustic state changes that conventional vibration sensors cannot detect.

[0042] Temperature signal monitoring utilizes a thermocouple array. The sampling frequency of the thermocouple array is set to 1Hz. Since heat conduction in the soil medium is a slowly evolving physical process, this low sampling frequency reduces data redundancy while reflecting the true thermal field evolution cycle. Due to environmental differences, there is a basic temperature difference between the fluid transported inside the pipe and the external soil medium. After the leaking fluid diffuses outward into the surrounding medium, it directly alters the local soil moisture content and specific heat capacity parameters, changing the original heat conduction path and thus forming a change in the local spatial thermal field distribution over time. The thermocouple array senses the temperature difference voltage at multiple spatial nodes based on the Seebeck effect at the contact surfaces of different metal conductors. A cold junction temperature measuring element is connected to the back end of the thermocouple array to perform cold junction compensation calculations, eliminating measurement errors caused by fluctuations in the ambient basic temperature. The system records the temperature data output by each thermocouple node within the array, acquiring the temperature gradient distribution along the radial and axial directions of the pipe. For the specific hardware circuit design and temperature compensation conversion calculations for thermocouple cold junction compensation, those skilled in the art can consult conventional sensor manuals for configuration based on measurement accuracy requirements. The circuit design is well-known in the field and will not be elaborated upon here.

[0043] The multi-dimensional data acquisition configuration employs a multi-channel parallel synchronous acquisition mechanism. The acquisition module internally features a system global clock source. This global clock source utilizes a temperature-controlled crystal oscillator to provide stable clock pulses. The global clock source distributes the reference clock signal to the corresponding sampling channels of each sensor via a synchronous communication bus. The analog-to-digital converters within each channel synchronously perform sampling operations based on the same trigger edge of this reference clock signal, such as the rising or falling edge. To ensure the accuracy of subsequent spatial positioning calculations based on time difference of arrival (TDOA), the clock synchronization accuracy is set to the microsecond level to reduce time difference measurement errors caused by sampling delays in each channel. This hardware-level clock synchronization configuration digitizes pressure, vibration, acoustic emission, and temperature signals on the same physical time scale, providing a common time reference for subsequent multi-source feature data stitching and TDOA spatial positioning calculations.

[0044] The sensor is installed on the pipeline, with a cable transmission distance between it and the acquisition module. For pressure sensors containing internal piezoresistive elements, the acquisition module uses a four-wire electrical connection path to acquire analog signals. The four-wire connection is configured with two independent current excitation lines and two independent voltage measurement lines. The current excitation lines provide a constant operating current to the sensor, and the voltage measurement lines are connected to the high-impedance input terminal of the acquisition module to extract the feedback voltage across the sensor. According to basic circuit principles, the high-impedance input terminal makes the current in the measurement loop approach zero, at which point almost no voltage drop occurs on the voltage measurement line. No operating current flows through the voltage measurement line loop, thus avoiding voltage drop interference caused by the parasitic resistance of the long transmission cable itself on the measured voltage value. The specific wiring structure and operational amplifier selection for the four-wire measurement circuit can be designed by those skilled in the art based on the pipeline distance on site; its circuit layout is well-known in the field and will not be elaborated here.

[0045] The sensor data from a multi-sensor network can be aggregated and transmitted through the acquisition module. Specifically, the acquisition module performs high-bit quantization conversion of weak signals. The physical field fluctuations caused by early pipeline leakage are relatively small, so the acquisition module is equipped with an analog-to-digital converter (ADC) chip with a resolution greater than or equal to 16 bits. Increasing the quantization bit depth of the ADC reduces the quantization interval, decreasing quantization errors during the conversion of analog voltage signals to discrete digital signals. Taking a pressure sensor with a range of 0–1 MPa paired with a 16-bit ADC as an example, its theoretical minimum resolvable pressure change reaches approximately 15 Pa. A refined quantization step is used to preserve the original form of weak acoustic emission elastic waves and small pressure drop fluctuations.

[0046] The acquisition module reads the converted digital values ​​from each channel using its built-in logic controller and performs hardware-level data alignment according to the set data frame format. The acquisition module concatenates and combines pressure, vibration, acoustic emission, and temperature data acquired within the same clock sampling period into the data payload area of ​​a single standard data packet. The acquisition module writes the clock sequence number of the sampling period as an absolute timestamp in the packet header area. The encapsulated data packets are sequentially stored in a data buffer queue and then transmitted to the edge gateway via a wired network interface, completing the parallel acquisition and physical encapsulation of raw data from the multi-dimensional heterogeneous sensors.

[0047] After receiving the multidimensional raw data sent by the acquisition module, the edge gateway calls the corresponding digital filter to perform filtering operations based on the frequency differences of the physical signals sensed by different types of sensors, so as to eliminate the interference of the on-site environment and separate the target frequency band that characterizes the leakage state.

[0048] For pressure signals, the edge gateway uses a finite impulse response (FIR) low-pass filter for processing. This FIR low-pass filter has a cutoff frequency set to 50Hz, employs a Hanning window function, and has a filter order of 64. In industrial environments, the AC power supply network introduces 50Hz power frequency electromagnetic interference, while the fluid within the pipe itself exhibits high-frequency random turbulence noise. Setting the cutoff frequency to 50Hz for the low-pass filter effectively suppresses high-frequency random noise and power frequency interference. The linear phase characteristic of the FIR filter ensures that the extracted small pressure drop waveform does not undergo phase distortion on the time axis.

[0049] For vibration signals, the edge gateway uses a cascaded second-order Butterworth bandpass filter for processing. The passband frequency range of this bandpass filter is set to 1kHz to 5kHz. The external environment of the pipeline typically contains low-frequency mechanical noise caused by surrounding vehicle traffic or routine machinery operation, with the frequency distribution of this noise concentrated below 1kHz. The structural resonance excited by the leakage water flow scouring the pipe wall and the surrounding medium is mainly distributed in the 1kHz to 5kHz frequency band. By using a bandpass filter with a passband frequency of 1kHz to 5kHz, and utilizing the maximum flatness characteristic of the Butterworth filter within the passband, it is possible to filter out low-frequency background environmental noise while retaining the resonant frequency components caused by water flow scouring, thus preventing energy distortion of the vibration waveform.

[0050] For acoustic emission signals, the edge gateway uses a high-pass filter for processing. The cutoff frequency of this high-pass filter is set to 15 kHz. Acoustic emission signals are used to characterize the frictional jet of high-pressure fluid passing through micropores and the transient elastic waves generated by the microscopic deformation of the pipe. The frequency distribution of elastic waves induced by these microscopic physical phenomena is higher than that of conventional macroscopic mechanical vibrations. Setting a high-pass cutoff frequency of 15 kHz can isolate low-frequency structural vibrations and environmental acoustic interference superimposed on the acoustic emission sensor, extracting the high-frequency transient elastic wave components and providing basic data for subsequent transient waveform envelope analysis.

[0051] For temperature signals, the edge gateway employs a moving average filtering algorithm. Thermal field changes in the soil medium are a slowly evolving physical process. During data acquisition, the temperature sensor array may experience random electrical glitches or high-frequency spikes due to external electromagnetic coupling. The moving average filtering algorithm calculates the arithmetic mean of temperature samples within a set data time window, using this average to replace the original sampling point at the center of the window. This data time window is set to a length of 5 to 10 seconds. Because the thermal field changes over a relatively long timescale, this time window length smooths out short-term glitches caused by electromagnetic coupling, while preventing the true gradual temperature change trend from being excessively smoothed out by an excessively long time window. This operation eliminates the interference of transient high-frequency abrupt changes on the temperature gradient difference calculation.

[0052] For the difference equation expressions, transfer function derivations, and discretization calculation processes of the filter weight coefficients corresponding to the above-mentioned digital filters, those skilled in the art can implement them by programming based on conventional digital signal processing theory. The underlying algorithm configuration is a well-known technology in this field and will not be elaborated here.

[0053] After the basic filtering operation is completed, the edge gateway performs composite denoising calculations on each signal to remove broadband random noise and environmental background interference that are unrelated to the physical characteristics of the leakage. The composite denoising calculation combines discrete wavelet transform and least mean square adaptive filtering algorithm.

[0054] The edge gateway extracts the filtered time-domain signal sequence and performs discrete wavelet decomposition. The db4 wavelet basis function is selected, as it possesses tight support and smoothness, corresponding physically to the transient waveforms caused by fluid leakage. The edge gateway sets the wavelet decomposition level to four, using a multi-scale pyramid algorithm to decompose the input one-dimensional time-domain signal into one low-frequency approximation coefficient component and four high-frequency detail coefficient components in different frequency bands. From a physical perspective, the energy of the real leakage signal in the wavelet domain is concentrated in a few large coefficients, while the energy of broadband white noise is evenly distributed across all wavelet coefficients with relatively small values. The high-frequency detail coefficient components contain local abrupt changes and broadband Gaussian white noise.

[0055] The edge gateway employs an adaptive soft thresholding algorithm to perform threshold quantization on four high-frequency detail coefficient components. The processing requires pre-calculating the adaptive threshold for each layer of coefficients. The core formula for calculating the adaptive threshold is as follows:

[0056] ;

[0057] in, Indicates an adaptive threshold. This represents an estimate of the noise standard deviation. This represents the total number of signal sampling points within the current processing window. It is also an estimate of the noise standard deviation. The result is calculated by taking the median absolute value of the first-layer high-frequency detail coefficients and dividing it by a constant of 0.6745. This calculation method utilizes the physical characteristic that the first layer of a high-frequency signal is mainly composed of high-frequency noise to obtain accurate noise distribution characteristics.

[0058] The soft threshold algorithm will select values ​​whose absolute values ​​are less than the adaptive threshold. The high-frequency detail coefficients are forced to zero, and their absolute values ​​are set to be greater than the adaptive threshold. The high-frequency detail coefficients shrink equally towards zero. The retained high-frequency detail coefficients correspond to the abrupt physical states caused by leakage. For the specific filtering differential calculations and the mathematical mapping process of the soft threshold function in the multi-scale tower algorithm, those skilled in the art can implement them through programming based on conventional digital signal processing theory. The underlying computational logic is well-known in the field and will not be elaborated upon here.

[0059] The edge gateway performs an inverse wavelet transform using the retained low-frequency approximation coefficient components and the high-frequency detail coefficient components after soft-threshold quantization to reconstruct a one-dimensional preliminary denoised signal in the time domain. This preliminary denoised signal eliminates broadband white Gaussian noise.

[0060] For persistent, slowly varying residual background noise in underground enclosed or highly damped environments, the edge gateway inputs the preliminary denoised signal to a minimum mean square adaptive filter (MINS). The MINS dynamically adjusts the tap weight coefficients through iterative optimization to approximate the statistical characteristics of the residual noise and separate it from the preliminary denoised signal. In the physical implementation logic, the edge gateway inputs the background signal collected by environmental reference sensors deployed in non-monitoring areas as the input reference noise signal vector to the filter. This noise signal is then weighted and summed using a finite impulse response network (FIR) within the filter to output an estimated noise signal. The edge gateway subtracts the estimated noise signal from the preliminary denoised signal to obtain an error signal used for feedback adjustment. The core formula for updating the filter tap weight coefficients is as follows:

[0061] ;

[0062] in, Indicates the first The filter weight vector of the next iteration Indicates the first The filter weight vector of the next iteration Indicates the step size factor. Indicates the first Error signal of the next iteration Indicates the first The input reference noise signal vector for the next iteration.

[0063] Step size factor The value range is set to 0.001 to 0.01. This range balances the convergence speed of weight updates with the minimum mean square error after the system reaches steady state. Setting this range prevents the step size factor from being affected. An excessively large filter can cause algorithmic divergence, while an excessively small filter can lead to system response lag. After the least mean square adaptive filter completes its iterations, the edge gateway outputs the physical signal after composite denoising processing.

[0064] After completing the composite denoising calculation, the edge gateway acquires physical signals including pressure, vibration, acoustic emission, and temperature data. These physical signals differ in dimensions and numerical magnitude. Due to the different working principles and units of measurement of various sensors, directly performing feature splicing and fusion calculations on data with different dimensions will result in physical components with larger absolute values ​​having a larger weight in subsequent neural network model calculations. This will reduce the weight of signals with smaller absolute values ​​but containing changes in the physical state of leakage, affecting the model's convergence. To eliminate the calculation bias caused by the difference in dimensions, the edge gateway uses a maximum-minimum normalization algorithm to uniformly map different types of physical signals to a set dimensionless numerical range, which is set to 0 to 1.

[0065] The edge gateway uses a sliding time window to extract a segment of the time-domain signal sequence as the current data processing window. It then extracts the boundary extrema of the signal sequence within this window using a traversal search algorithm and performs a linear transformation. At the physical computation level, to prevent a constant signal within the current processing window (i.e., when the maximum and minimum values ​​are equal, triggering a division-by-zero crash at the program's lower level), the edge gateway introduces a minimal positive real number as a protection mechanism in the denominator of the normalization calculation. The core formula for maximum-minimum normalization is as follows:

[0066] ;

[0067] in, This represents the normalized signal sequence value. This represents the current sampling point value of the input signal sequence. This represents the minimum sample value of the signal sequence within the current data processing window. This represents the maximum sample value of the signal sequence within the current data processing window. This represents the smallest positive real number used to prevent division by zero errors, and its value is set to 10. -8 .

[0068] During physical implementation, the sliding time window length set for the edge gateway is consistent with the time period of the standard data packets in the multi-channel parallel synchronous acquisition mechanism. This parameter setting ensures that each normalization calculation corresponds to multi-dimensional sensor data within the same physical time span, preventing data timing shifts caused by mismatched time window lengths. Simultaneously, the use of a dynamic extreme value update method based on local time windows can adapt to the slow drift of basic environmental parameters during long-term pipeline operation.

[0069] After maximum and minimum value normalization, each signal is transformed into a dimensionless preprocessed signal with a unified numerical range, while retaining the original waveform envelope contour and relative amplitude variation trend. The normalized data serves as the standardized input for the subsequent feature extraction module. For the specific data buffer allocation and matrix mapping calculation process in the normalization process, those skilled in the art can write conventional programs based on the storage architecture of the microprocessor inside the edge gateway. The underlying matrix operation logic is well-known in the field and will not be elaborated here.

[0070] The first processor receives the normalized pressure signal sequence and divides the data into segments based on a set time window, then performs time-domain statistical feature extraction. The length of this time window is consistent with the standard data packet time period in the multi-channel parallel synchronous acquisition mechanism, used to match the response time of fluid dynamics in the pressure pipeline network. The time-domain feature extraction is based on the amplitude distribution of the discrete signal, used to characterize the fluid dynamic changes caused by pipeline leakage.

[0071] The first processor calculates the mean parameter of the pressure signal. Local leakage in the pipeline causes fluid to flow outward, disrupting the original fluid pressure balance, which manifests as a decrease in the base water pressure in the sensor installation area. The mean parameter is obtained by calculating the arithmetic mean of all normalized sampling point values ​​within the current data processing window, and is used to reflect the evolution trend of the base pressure drop.

[0072] The first processor extracts the variance parameter of the pressure signal. When fluid is ejected outward through a ruptured hole in the pipe wall, the geometric discontinuities at the hole's edge induce local turbulence and vortices, causing dynamic fluctuations to superimpose on the static water pressure inside the pipe. The variance parameter measures the dispersion of these dynamic fluctuations. The core formula for variance calculation is as follows:

[0073] ;

[0074] in, This represents the variance of the pressure signal. This indicates the total number of signal sampling points within the current processing window. This represents the first value in the normalized signal sequence. Values ​​of each sampling point This represents the arithmetic mean of the normalized signal sequence values ​​within the current data processing window.

[0075] The first processor extracts the peak-to-peak value and kurtosis parameters of the pressure signal. The peak-to-peak value is the difference between the maximum and minimum extreme values ​​of the pressure signal within the current data processing window. Its physical significance lies in quantifying the transient pressure range caused by the instantaneous pipe rupture and internal water hammer effect. The kurtosis parameter characterizes the steepness of the pressure signal amplitude probability density distribution curve relative to a normal distribution. Mathematically, the kurtosis parameter is calculated from the ratio of the fourth central moment to the square of the variance of the pressure signal sequence. When the pipeline is in a stable, leak-free operating state, the background pressure fluctuations approximately follow a normal distribution; when a leak occurs and a transient shock wave is triggered, the tail of the signal amplitude distribution curve thickens. Calculating the kurtosis parameter can identify the pulse impact components contained in the pressure signal sequence.

[0076] The first processor concatenates and combines the extracted single parameter values ​​such as mean, variance, peak-to-peak value, and kurtosis according to a predetermined order to construct a one-dimensional stress time-domain feature vector. This feature vector constitutes a subset of stress data in the multi-dimensional feature space. The specific memory addressing logic and low-level calculation process of the time-domain statistical features within the microprocessor can be implemented by those skilled in the art using conventional embedded system development manuals. The underlying operations are well-known in the field and will not be elaborated upon here.

[0077] The first processor receives the normalized vibration signal sequence within the current data processing window and performs frequency domain and time-frequency domain feature extraction operations based on the frequency distribution of the signal sequence. The vibration signal is used to characterize the structural resonance state induced by fluid scouring of the pipe wall, which has a regular energy distribution characteristic in the frequency domain.

[0078] The first processor uses a Fast Fourier Transform (FFT) to convert the time-domain vibration signal sequence to the frequency domain, obtaining the corresponding power spectrum. Based on the power spectrum data, the edge gateway calculates the centroid frequency of the vibration signal. The centroid frequency is calculated by summing the products of each frequency component and its power amplitude, divided by the sum of the power amplitudes. Its physical meaning lies in quantifying the center position of the spectral energy distribution. When a pipe leaks, the resonance caused by the water flow impact will shift the frequency band where energy is concentrated. Calculating the centroid frequency can capture this frequency band shift trend. The core formula for calculating the centroid frequency is as follows:

[0079] ;

[0080] in, Indicates the centroid frequency. This represents the total number of frequency lines in the power spectrum. Indicates the first The frequency values ​​corresponding to each frequency spectral line Indicates the first The power spectral amplitude corresponding to each frequency spectral line.

[0081] Because pipeline vibration signals possess non-stationary and nonlinear physical properties, a single frequency domain analysis cannot reflect the energy variation over time. Therefore, the edge gateway employs wavelet packet transform for time-frequency feature extraction. At the algorithm configuration level, the edge gateway selects the db4 wavelet basis, consistent with the preceding denoising step, as the decomposition basis function to match the transient vibration waveform characteristics induced by water flow impact. Wavelet packet transform extends the orthogonal decomposition frequency band of the vibration signal to the high-frequency band, obtaining a complete time-frequency distribution map. The edge gateway sets the wavelet packet decomposition level to 3 layers, decomposing the vibration signal into 8 independent frequency band nodes. The edge gateway calculates the total signal energy within each frequency band node and the proportion of energy from a single node to the total energy. For the specific implementation of signal energy calculation, the first... The signal energy of each frequency band node is calculated by the sum of the squares of the absolute values ​​of all signal coefficients within that node. The total energy is the sum of the signal energies of all eight frequency band nodes. The proportion of a single node's energy to the total energy is the quotient of that node's signal energy divided by the total energy. Based on this energy proportion, the edge gateway calculates the wavelet packet energy entropy. Wavelet packet energy entropy measures the degree of disorder in the energy distribution of vibration signals within each frequency band. When the pipeline is leak-free, the vibration energy distribution in the background environment is relatively dispersed, resulting in a higher energy entropy value. When leakage occurs, the vibration energy is concentrated in the target frequency band where water flow erodes the pipe wall, causing resonance, leading to a decrease in the disorder of the energy distribution and a corresponding decrease in the energy entropy value. The core formula for calculating wavelet packet energy entropy is as follows:

[0082] ;

[0083] in, This represents the wavelet packet energy entropy. This represents the total number of frequency band nodes in the wavelet packet decomposition. Indicates the first The proportion of energy of each frequency band node to the total energy.

[0084] The first processor concatenates the extracted centroid frequency and wavelet packet energy entropy in a predetermined order to construct a one-dimensional vibration feature vector. This feature vector constitutes a subset of vibration data in the multi-dimensional feature space. The underlying matrix multiplication and addition operations and memory allocation logic of the Fast Fourier Transform and Wavelet Packet Transform can be implemented by those skilled in the art based on conventional digital signal processing theory and microprocessor manuals. The underlying algorithm configuration is well-known in the field and will not be elaborated upon here.

[0085] The first processor receives the normalized acoustic emission signal sequence within the current data processing window. Acoustic emission signals are used to characterize the frictional jet of high-pressure fluid passing through the microscopic pores of the pipe wall and the transient elastic waves generated by the microscopic deformation of the pipe material. The signals generated by these physical phenomena exhibit a sudden and energy-concentrated wave packet shape.

[0086] The first processor uses Hilbert transform to perform envelope detection on the acoustic emission signal sequence to obtain the envelope curve of the acoustic emission signal. The Hilbert transform constructs an analytic signal from the original signal through quadrature phase shifting and calculates the magnitude of this analytic signal. The physical principle of this process lies in stripping away the carrier oscillation component of the high-frequency elastic wave and extracting the low-frequency envelope profile that reflects the transient trend of energy changes.

[0087] The first processor sets a dynamic trigger threshold based on the acquired envelope curve to identify transient acoustic emission events. Since the background acoustic state of the pipeline installation environment slowly drifts with the external environment, using a fixed value as the trigger reference could lead to missed weak signals or misjudgments of environmental noise. The first processor multiplies the arithmetic mean of all sampling points on the envelope curve within the current data processing window by the set trigger coefficient to obtain the dynamic trigger threshold. The trigger coefficient is set to a range of 1.5 to 3.0. This range is set to achieve a balance between filtering out stable background noise and retaining the initial wave packets of weak leakage.

[0088] The first processor uses a dynamic trigger threshold to determine the truncation of the envelope curve. When the value of the envelope curve increases and crosses the dynamic trigger threshold, the first processor marks the sampling number corresponding to the crossing point as the event start point. Subsequently, when the value of the envelope curve decreases and crosses the dynamic trigger threshold again, to prevent the attenuation tail of the acoustic emission signal from oscillating near the threshold and causing the same event to be misjudged as multiple independent fragment wavelet packets, a set event locking time window is activated. The length of the event locking time window is set to 300 microseconds. If the value of the envelope curve does not rise again above the dynamic trigger threshold within the event locking time window, the sampling number corresponding to the fallback point is marked as the event end point. The envelope curve data segment from the event start point to the event end point constitutes an independent acoustic emission transient event.

[0089] The first processor extracts feature parameters from the identified acoustic emission transient events. It counts the total number of acoustic emission transient events identified within the current data processing window and uses this count as the impact count parameter. The impact count parameter measures the frequency of occurrence of the microporous friction jet on the time axis. For a single acoustic emission transient event, it calculates the cumulative sum of the values ​​of all envelope curve sampling points within the corresponding data segment and uses this sum as the single event energy parameter. The core formula for calculating the single event energy is as follows:

[0090] ;

[0091] in, Represents the energy parameter of a single event. The sampling number indicating the starting point of the event. The sampling number indicating the end point of the event. Indicates the first element in the envelope curve The values ​​of each sampling point.

[0092] The first processor extracts the maximum envelope value within a single acoustic emission transient event data segment as the single-event peak amplitude parameter. The single-event peak amplitude parameter is used to characterize the maximum instantaneous energy level of the micro-deformation or jet release. To construct a unified feature dimension matching the current data processing window, the arithmetic mean of the energy parameters of all identified single events within the current data processing window is calculated to obtain the average event energy parameter; simultaneously, the arithmetic mean of the peak amplitude parameters of all single events within the current data processing window is calculated to obtain the average peak amplitude parameter.

[0093] The first processor concatenates and combines the impact count parameter, average event energy parameter, and average peak amplitude parameter in a predetermined order to construct a one-dimensional acoustic emission feature vector. This feature vector constitutes a subset of acoustic emission data in the multi-dimensional feature space. The discrete convolution operation of the Hilbert transform and the internal register comparison logic for event threshold determination can be implemented by those skilled in the art using conventional digital signal processing theory. The underlying mathematical operations are well-known in the field and will not be elaborated upon here.

[0094] The first processor acquires the normalized temperature signal sequence from multiple temperature sensor nodes distributed along the pipeline space within the current data processing window. Fluid leakage into the surrounding soil or insulation layer causes localized heat exchange, disrupting the original thermal equilibrium of the medium. Due to the hysteresis and spatial locality of heat transfer, leakage creates an abnormal temperature distribution gradient along the pipeline's axial spatial dimension. Based on this physical principle, the spatial characteristics of the temperature field are extracted.

[0095] The first processor extracts the maximum spatial temperature difference parameter. This parameter measures the extreme thermal difference between the leakage center area and the unaffected area. The arithmetic mean of all normalized temperature sampling points for each temperature sensor node within the current data processing window is calculated and used as the baseline temperature value for that node. By traversing and comparing the baseline temperature values ​​of all spatial nodes, the maximum and minimum values ​​are extracted. The difference between the maximum and minimum values ​​is then calculated, and this difference is the maximum spatial temperature difference parameter.

[0096] The first processor extracts the spatial temperature gradient parameter. This parameter quantifies the uneven distribution of the thermal field in the surrounding medium. When leakage occurs at a single node, causing a localized temperature anomaly, the temperature difference between adjacent nodes increases. The absolute value of the difference in baseline temperature values ​​between adjacent temperature sensor nodes is calculated and divided by the physical distance between the nodes. The core formula for calculating the spatial temperature gradient is as follows:

[0097] ;

[0098] in, Represents the spatial temperature gradient parameter. This indicates the total number of temperature sensor nodes. Indicates the spatial index of the temperature sensor node. Indicates the first The base temperature value of each temperature sensor node. Indicates the first The base temperature value of each temperature sensor node. This represents the physical distance between adjacent temperature sensor nodes. The physical distance between adjacent temperature sensor nodes is measured and obtained by the pipeline monitoring system during the construction and deployment phase, and is pre-stored in the non-volatile memory of the edge gateway for later retrieval.

[0099] The first processor extracts the spatial coefficient of variation parameter. This parameter measures the non-uniformity of the overall temperature field distribution along the pipeline. When the pipeline is leak-free and operating normally, the ambient temperature influence on each node is basically the same, resulting in a small spatial coefficient of variation. However, when local leakage occurs, forming a concentrated thermal anomaly, the spatial coefficient of variation increases. The standard deviation of the baseline temperature values ​​for all nodes is calculated, and this standard deviation is divided by the arithmetic mean of all node baseline temperature values ​​to obtain the spatial coefficient of variation parameter. At the physical calculation level, since the node baseline temperature values ​​are derived from normalized values, their arithmetic mean may be zero under isothermal extreme conditions. To prevent division errors at the program's underlying level from causing division-to-zero crashes, a minimal positive real number is introduced as a protection mechanism in the denominator of the spatial coefficient of variation parameter calculation. This minimal positive real number is set to 10. -8 .

[0100] The first processor concatenates and combines the spatial maximum temperature difference parameter, spatial temperature gradient parameter, and spatial coefficient of variation parameter in a predetermined order to construct a one-dimensional spatial feature vector of the temperature field. This feature vector constitutes a subset of temperature data in the multi-dimensional feature space. Then, the previously acquired pressure time-domain feature vector, vibration feature vector, and acoustic emission feature vector are concatenated and concatenated with this spatial feature vector of the temperature field to generate a comprehensive fusion feature vector characterizing the multi-dimensional physical states of fluid dynamics and thermodynamics within the current data processing window, providing complete input data for the subsequent leakage identification model. The spatial matrix traversal logic and underlying division instruction operation process of the temperature field features can be implemented by those skilled in the art using conventional microprocessor instruction sets. The underlying mathematical operations are well-known in the field and will not be elaborated upon here.

[0101] After calculating the multi-dimensional physical parameters, the dispersed heterogeneous feature data is obtained. To construct a unified data structure suitable for subsequent dimensionality reduction and model recognition algorithms, the second processor performs a spatiotemporal dimension-based concatenation operation on the pressure time-domain feature vector, vibration feature vector, acoustic emission feature vector, and temperature field spatial-domain feature vector.

[0102] The second processor performs time-dimension alignment. There are physical differences in the propagation and conduction speeds of fluid dynamics and thermodynamics in the pipe medium. Without time-base constraints, directly combining features extracted from different time sequences will lead to misalignment in the representation of the leakage's physical state. The physical principle of time-base constraints lies in the fact that the propagation speeds of pressure transient pulses and acoustic emission waves generated by fluid jets are on the order of kilometers per second, while the thermal conduction of the temperature field to form a stable temperature gradient often has a long time lag. By using a unified time tag for hard data truncation and combination, transient dynamic parameters and steady-state thermodynamic parameters within the same absolute physical time interval can be locked in the same frame of data. The second processor reads the global timestamp tags carried by each feature vector in the header file of the internal buffer. Using the system clock cycle of the current data processing window as a reference, the global timestamp tags carried by the pressure time-domain feature vector, vibration feature vector, acoustic emission feature vector, and temperature field spatial-domain feature vector are verified. If the timestamps of the above four sets of feature vectors are all within the span of the same current data processing window, time alignment is determined, and the four sets of feature vectors are loaded into the cascaded register group in parallel. If any of the four sets of feature vectors contains a timestamp exceeding the current data processing window span or a missing data header file, the temporal alignment is deemed to have failed. In the event of alignment failure, to maintain a fixed width of the spliced ​​feature dimension, the historical valid feature data of the corresponding physical parameter within the previous adjacent data processing window is retrieved and zero-order preserved padding replacement is performed to ensure the integrity of the input data for subsequent dimensionality reduction models.

[0103] Perform a structured sort along the spatial dimension. Retrieve the spatial index of the sensor physical topology configuration table pre-stored in non-volatile memory to obtain the spatial index of the data acquisition nodes distributed along the pipeline axis. Following the ascending order of the spatial index, cluster the pressure time-domain feature vectors, vibration feature vectors, and acoustic emission feature vectors belonging to the same monitoring section coordinate into local physical parameter sets. Subsequently, the temperature field spatial feature vector reflecting the heat conduction state of the entire pipeline is appended as a global physical parameter to the region following this local physical parameter set.

[0104] By performing a cascaded mapping operation on the matrix array through the data bus, the local physical parameter set is merged with the global physical parameters to generate a high-dimensional data array across dimensions. This array is the comprehensive fusion feature vector. The core formula for concatenating the comprehensive fusion feature vector is as follows:

[0105] ;

[0106] in, This represents the integrated feature vector. Represents the time-domain eigenvector of pressure. Represents the vibration eigenvector. Represents the acoustic emission eigenvector. This represents the spatial eigenvectors of the temperature field. The concatenation operator for multidimensional vectors.

[0107] In this cascaded architecture, multi-source heterogeneous scalar parameters are integrated into a one-dimensional feature row vector with a physical spatiotemporal correspondence. This integrated feature vector eliminates the data length differences caused by different sampling frequencies in the original data, forming a standardized input format with a fixed dimensional width. The integrated feature vector generated after concatenation is transmitted to the analysis platform as the input benchmark data for subsequent structured dimensionality reduction models. Regarding the address pointer jump allocation and direct memory access and movement logic of the feature vector within the edge gateway's static random access memory, those skilled in the art can write conventional instructions based on the conventional microprocessor underlying memory management mechanism. The underlying addressing rules and data movement mechanisms are well-known technologies in this field and will not be elaborated upon here.

[0108] The analysis platform receives the integrated feature vector transmitted from the edge gateway. Before inputting this vector into the subsequent dimensionality reduction and recognition model, the analysis platform performs batch normalization processing on the integrated feature vector.

[0109] In the computation of deep neural networks, as the network depth increases, the probability distribution of the input data in intermediate layers continuously changes, which can easily lead to gradient vanishing or computational divergence during model parameter updates. Since the integrated feature vector contains spatiotemporal multidimensional features calculated through different physical mechanisms, the physical parameters of different dimensions still differ in absolute value and fluctuation range. The physical significance of batch normalization lies in mapping the input integrated feature vector to a probability distribution space with a uniform mean and variance, eliminating the internal covariate bias caused by splicing multi-source heterogeneous parameters, and enabling subsequent network models to extract leakage features from the pipeline network.

[0110] Since the analysis platform is currently performing inference and analysis tasks for online leakage monitoring, the input data is often a single sample or a small batch of samples. Calculating the arithmetic mean and variance of the current batch of samples in real time could easily lead to statistical distribution distortion due to insufficient sample size, or even result in a calculation error where the variance is zero when only a single sample is input. Therefore, the analysis platform reads the global moving mean and global moving variance accumulated and fixed during the model training phase from the pre-trained deep neural network model parameter weight file. The analysis platform uses the read global moving mean as the batch mean and the read global moving variance as the batch variance. The analysis platform then uses the acquired batch mean and batch variance to standardize the currently input integrated feature vector, making it conform to a data distribution with an approximate mean of 0 and a variance of 1.

[0111] During the standardization mapping process, the standardized data alters the fluid dynamics and thermodynamic distribution relationships inherent in the original integrated feature vector, potentially losing some leakage characterization information. The analysis platform performs linear reconstruction on the standardized feature data by introducing scaling and translation parameters to restore the data's feature expressiveness. The analysis platform integrates the above standardization and linear reconstruction steps into a single matrix operation, the core formula of which is as follows:

[0112] ;

[0113] in, This represents the feature vector after batch normalization. This represents the integrated feature vector. This represents the batch average. Indicates batch variance. This represents the smallest positive real number used to prevent division by zero errors, and its value is set to 10. -8 , This represents the scaling parameter. This represents the translation parameter.

[0114] At the underlying operational protection level of the formula, a minimal positive real number is introduced inside the square root sign of the denominator. As a protection mechanism, it prevents the microprocessor from crashing during division-by-zero operations due to the constant environmental data causing the historical variance to approach zero. The scaling parameter in the formula... With translation parameters The analysis platform reads and calls the parameter weights from the pre-trained deep neural network model.

[0115] After batch normalization, the integrated feature vectors are converted into standardized input data with stable distribution and preserved leakage physical characteristics. For the underlying matrix parallel accumulation mechanism and the logic for arranging and calling tensor data within the microprocessor, those skilled in the art can programmatically schedule it according to the implementation methods of underlying operators in conventional deep learning frameworks. The underlying tensor multiplication and addition operation logic is well-known in the field and will not be elaborated upon here.

[0116] The analysis platform constructs a hybrid deep neural network consisting of a cascaded one-dimensional convolutional neural network and a long short-term memory network to identify and determine the status of pipeline leaks. The platform uses batch-normalized feature vectors as the underlying input data, feeding them into the input layer of this hybrid deep neural network to initiate the forward propagation computation process.

[0117] The analysis platform utilizes a one-dimensional convolutional neural network to extract local cross-domain correlation features from heterogeneous data. The integrated feature vector sequentially arranges physical parameters of different dimensions, such as pressure, vibration, acoustic emission, and temperature, with fluid dynamics and thermodynamic coupling between adjacent feature elements. The one-dimensional convolutional neural network performs sliding matrix multiplication and addition operations on the batch-normalized feature vector using a convolutional kernel of a set length, mapping the correlation information of different physical dimensions to a higher-dimensional feature map. The core formula for one-dimensional convolution calculation is as follows:

[0118] ;

[0119] in, The output of a one-dimensional convolutional layer is shown. Local feature mapping values, Represents the sliding sequence index of the local feature map. This indicates a modified linear unit activation function. This indicates the length of the one-dimensional convolution kernel. Represents the first in the convolution kernel One weight parameter, This represents the feature vector after batch normalization. Indicates the spatial index of the temperature sensor node. This represents the bias parameters of the one-dimensional convolutional layer. At the network structure parameter setting level, the analysis platform specifies the length of the one-dimensional convolutional kernel. Set to 3, and index the spatial sequence number of the temperature sensor node. Set to 1 to iterate through adjacent physical parameter features one by one.

[0120] After completing the convolution operation, the analysis platform performs max pooling dimensionality reduction on the output feature mapping matrix. Max pooling extracts the maximum value within the local receptive field, filters out background feature interference, retains energy fluctuation information representing the leakage state, and compresses the dimensionality of the data matrix, reducing the computational load for subsequent time-series processing.

[0121] The analysis platform inputs the pooled and dimensionality-reduced feature sequence into the Long Short-Term Memory (LSTM) network layer. Although the input fused feature vector is static slice data in a single computation, the feature sequence generated by one-dimensional convolution and pooling operations contains structural dependencies in spatial dimensions. The LTM network, through its internal gating mechanism, controls the retention of historical state information and the writing of current input information, used to parse the spatial structural dependencies in the feature sequence. The core formula for updating the cell state of the LTM network is as follows:

[0122] ;

[0123] in, This indicates the current state of a cell in the Long Short-Term Memory network. This represents the output vector of the forget gate. The matrix element-wise multiplication operator is used to represent the matrix multiplication operator. This indicates the state of the long short-term memory network cells in the previous step. This represents the output vector of the input gate. This represents the candidate cell state vector for the current step. The output vector of the forget gate, the output vector of the input gate, and the candidate cell state vector for the current step in the above formula are all generated by the Long Short-Term Memory network after receiving the input feature sequence elements of the current step and the hidden layer state vector of the previous step, through internal linear mapping and a Sigmoid or hyperbolic tangent activation function.

[0124] As the forward propagation process progresses, the analysis platform obtains the hidden layer output vector of the last sequence step of the Long Short-Term Memory (LSTM) network and flattens it before passing it to the fully connected layer. The fully connected layer maps the high-dimensional abstract features to classification nodes equal to the preset number of leakage categories using matrix multiplication. The pipeline monitoring system sets the preset leakage state categories to four: normal operation, micropore leakage, moderate crack leakage, and severe pipe wall rupture. Therefore, the preset total number of leakage state categories, NCNC, is set to 4. The analysis platform calls the normalized exponential function to probabilistically map the output values ​​of the classification nodes, calculating the predicted probability of the current monitoring section being in each leakage state. The core formula for calculating the predicted probability is as follows:

[0125] ;

[0126] in, This indicates that the current monitoring segment belongs to the first... The predicted probability of various leakage states. This indicates that the fully connected layer corresponds to the first... Output values ​​for each category This indicates the total number of preset leakage status categories. This indicates that the fully connected layer corresponds to the first... Output values ​​for each category.

[0127] The analysis platform extracts the state label corresponding to the maximum predicted probability based on the calculated probability distribution set, and uses it as the final leakage diagnosis result for the current pipeline monitoring section. For the multiplication and addition operations of the underlying neurons in one-dimensional convolutional neural networks and long short-term memory networks, the boundary processing of activation functions, and the pointer offset scheduling of multi-dimensional tensors in the microprocessor's internal cache, those skilled in the art can program and deploy them according to conventional deep learning inference frameworks. The underlying graph computation unfolding logic is well-known technology in this field and will not be elaborated upon here.

[0128] The analysis platform uses a set of probability distributions output by a hybrid deep neural network to determine the leakage level and implement system response. To improve the reliability of the determination and prevent false alarms caused by environmental noise interference, the analysis platform performs a confidence threshold check before extracting the state label corresponding to the maximum predicted probability.

[0129] The analysis platform compares the maximum predicted probability value in the probability distribution set with a set confidence threshold. The confidence threshold is set to range from 0.75 to 0.85. If the maximum predicted probability value is greater than or equal to the confidence threshold, the analysis platform accepts the category index corresponding to the maximum value and outputs it as the leakage status label for the current monitoring section. The core logic of category determination is as follows:

[0130] ;

[0131] in, This indicates the output leakage status label. This indicates that the current monitoring segment belongs to the first... The predicted probability of various leakage states. This indicates the total number of preset leakage status categories, with a value of 4. This is an operator that finds the index of the argument that maximizes the subsequent function.

[0132] If the maximum predicted probability value is less than the confidence threshold, the analysis platform determines that the current data is in a critical state and marks the output result as pending. To address the issue of low confidence caused by incomplete feature extraction from a single data collection, the analysis platform initiates a moving average decision mechanism. The analysis platform caches the probability distribution set of the current data processing window and continuously acquires the probability distribution sets of the next two adjacent data processing windows, forming a moving decision sequence with a total length of three windows. The analysis platform calculates the arithmetic mean of the predicted probabilities of the same category within these three data processing windows, generating a smoothed probability distribution set. Subsequently, the analysis platform extracts the maximum smoothed predicted probability value from the smoothed probability distribution set again and compares it with a set secondary decision threshold. The secondary decision threshold is set to 0.60. If the maximum smoothed predicted probability value is greater than or equal to the secondary decision threshold, the analysis platform outputs the status label corresponding to the maximum smoothed predicted probability value; if the maximum smoothed predicted probability value is still less than the secondary decision threshold, the analysis platform forces the output of the leakage status label of the current monitoring section as normal operation, clears the current cache, and restores the initial single-window decision logic.

[0133] The analysis platform executes a tiered response strategy based on the identified leakage status labels. The four leakage status categories described above correspond to different action logics:

[0134] When the leakage status label corresponds to the normal operating status, it indicates that the pipeline medium is in a closed transport condition. The analysis platform stores the current status data in the local non-volatile storage medium to maintain the regular data acquisition cycle.

[0135] When the leakage status label corresponds to a micropore leakage condition, it indicates that the pipe wall has deformed or that there is a pinhole, and fluid is leaking out in a small amount. The analysis platform generates a level one early warning message, which is sent to the host computer control center via industrial Ethernet, containing a message with a timestamp and spatial node location, prompting maintenance personnel to perform planned inspections.

[0136] When the leakage status label corresponds to a medium-sized crack leakage state, it indicates localized cracking of the pipe material and a deviation in fluid dynamics and thermodynamic parameters. The analysis platform generates a level-two alarm command, sending an audible and visual alarm drive signal to the control center's operation panel; simultaneously, it sends a configuration command to the edge gateway of the corresponding node via the underlying data bus, increasing the sensor's data sampling frequency to twice the original configuration parameters to obtain high-resolution pipe network data.

[0137] When the leakage status label corresponds to a severe pipe wall rupture, it indicates structural failure in the pipeline network and a risk of fluid leakage. The analysis platform triggers a three-level emergency interlock mechanism, directly sending hard-wired interlock signals to the programmable logic controllers deployed at both ends of the target pipe section, driving the emergency shut-off valves to perform a shutdown operation, thus blocking the physical transport of fluid.

[0138] The hardware driving mechanisms for the floating-point comparison logic branches inside the microprocessor, the network socket packet transmission process of alarm messages, and the relay latching output circuit can be programmed and implemented by those skilled in the art according to conventional industrial control system standards. The underlying network communication and electrical control mechanisms are well-known technologies in this field and will not be elaborated here.

[0139] During system deployment or when the physical environment of the pipeline network changes, the analysis platform performs closed-loop training on the hybrid deep neural network to establish the weight and bias parameters within the one-dimensional convolutional neural network and the long short-term memory network.

[0140] The analysis platform retrieves accumulated integrated feature vectors from local non-volatile storage media during historical operation and, combined with manual inspection and maintenance records, assigns corresponding real leakage state labels to these feature vectors, thus constructing a structured training dataset. The real leakage state labels correspond to four preset categories: normal operation, micropore leakage, moderate crack leakage, and severe pipe wall rupture. The analysis platform inputs the training dataset into a hybrid deep neural network in batches, calculating the predicted probability distribution of each sample using a forward propagation mechanism.

[0141] The analysis platform calculates the deviation between the network's predicted output and the true label. Since leakage status identification is a multi-class classification task, the platform uses the cross-entropy loss function to quantify this deviation. In the mathematical mechanism of classification models, cross-entropy loss amplifies the numerical penalty for classification errors, giving the gradient descent process greater corrective momentum in the early stages of training. The cross-entropy loss function measures the difference between the predicted probability distribution and the true probability distribution by taking the logarithm of the predicted probabilities of the true label category and summing them. The core formula for calculating cross-entropy loss is as follows:

[0142] ;

[0143] in, This represents the cross-entropy loss value. This represents the total number of samples in the current training batch. This represents the sample index within the current training batch. This indicates the total number of preset leakage status categories, with a value of 4. Show the first The true label of the sample in the th... One-hot encoded values ​​for each category, The first term represents the output of the hybrid deep neural network. The sample belongs to the first The predicted probabilities of each category, This represents the natural logarithm operator.

[0144] Based on the calculated cross-entropy loss value, the analysis platform initiates the backpropagation calculation process. Following the chain rule, the platform calculates the gradient of the cross-entropy loss value with respect to the weight and bias parameters of each network node, layer by layer from the output layer to the input layer. The calculated gradient matrix reflects the direction and magnitude of the current network parameters deviating from the optimal solution.

[0145] The analysis platform employs an adaptive moment estimation (IME) algorithm to update network parameters. Compared to the basic stochastic gradient descent algorithm, IEM dynamically adjusts the update step size for each network parameter by calculating the exponential moving average of the gradient, thereby preventing the model parameters from getting trapped in local optima. The core formula for parameter updates is as follows:

[0146] ;

[0147] in, This indicates the updated network parameters. Indicates the network parameters for the current step. Represents the global learning rate. This represents the first-order momentum estimate after bias correction. This represents the second-order momentum estimate after bias correction. This represents the smallest positive real number used to prevent division by zero errors, and its value is set to 10. -8 In the underlying matrix derivation logic, the bias-corrected first-order momentum estimate is obtained by calculating and correcting the historical gradient data and the first-order momentum decay coefficient using an exponential moving average; the bias-corrected second-order momentum estimate is obtained by calculating and correcting the historical gradient squared data and the second-order momentum decay coefficient using an exponential moving average. In the system initialization configuration, the analysis platform sets the first-order momentum decay coefficient to 0.9 and the second-order momentum decay coefficient to 0.999.

[0148] To ensure the network parameters converge to their optimal point in the later stages of training, the analysis platform introduces an adaptive optimization strategy of learning rate decay during parameter updates. The analysis platform uses the global learning rate... The initial value was set to 0.001. As the training iteration cycle increased, the analysis platform gradually reduced the global learning rate according to the preset decay coefficient. The preset decay coefficient is set to 0.95. This strategy achieves a wide search of the parameter space in the early stages of training with a large learning rate, and prevents parameters from oscillating in the minimum region in the later stages of training with a smaller learning rate. When the analysis platform detects that the cross-entropy loss value converges to within the set lower limit of error, or the training process reaches the set maximum number of iterations, the closed-loop training is determined to be complete. The system sets the lower limit of error to 0.01 and the maximum number of iterations to 1000. The analysis platform saves the finally converged network parameters and reads and calls them in the actual online monitoring and inference process. For the chain multiplication operation of partial derivative matrices, the underlying iterative accumulation logic of momentum estimation, and the memory-resident scheduling of loss function values ​​in the backpropagation process, those skilled in the art can programmatically implement them according to conventional deep learning backpropagation algorithms. The underlying calculus operations and gradient descent mechanism are well-known technologies in this field and will not be elaborated here.

[0149] When the analysis platform determines that the leakage status is classified as micropore leakage, moderate crack leakage, or severe pipe wall rupture, the platform triggers a leakage location procedure based on multi-source information collaboration. The analysis platform performs an initial calculation of the distance to the leakage point based on time-of-flight analysis principles.

[0150] When fluid overflows from a pipe rupture, it causes a local pressure drop within the pipe, creating a negative pressure wave. This negative pressure wave propagates along the fluid medium within the pipe towards the upstream and downstream sides of the leak point. The analysis platform retrieves the pressure time-domain characteristic sequences synchronously collected by the upstream and downstream edge gateways of the target pipe section, using them as the data source for time-of-flight analysis.

[0151] Due to background mechanical noise and fluid turbulence interference in industrial environments, relying on the threshold exceeding time of a single signal to determine the wavefront arrival time is prone to judgment bias. The analysis platform employs a cross-correlation algorithm to perform time alignment analysis on the upstream and downstream pressure time-domain feature sequences. Cross-correlation measures the waveform similarity of two discrete-time signals at different time-shift scales. The analysis platform keeps the upstream pressure time-domain feature sequence fixed and slides the downstream pressure time-domain feature sequence point-by-point along the time axis, calculating the cross-correlation function sequence of the two sequences. It then iterates through this cross-correlation function sequence and finds the shift time point corresponding to the maximum peak. This shift time point reflects the relative delay between the two signals in the time domain. Based on this shift time point, the analysis platform calculates the absolute time of the negative pressure wave signal received by the upstream acquisition node minus the absolute time of the negative pressure wave signal received by the downstream acquisition node, using this value as the time difference.

[0152] Based on the acquired time difference, the analysis platform combines the spatial topology parameters of the sensor nodes with the physical properties of the medium to calculate the one-dimensional spatial coordinates of the leak point. The physical principle of negative pressure wave localization lies in the fact that the negative pressure wave generated when fluid overflows from a pipe rupture propagates to both sides at a fixed speed. Due to the difference in physical distance between the leak point and the upstream and downstream sensors, there is a time lag in the wavefront reaching both sensors. By calculating this time delay, the spatial distance can be derived in reverse using algebraic equations. The core formula for distance measurement based on time difference analysis is as follows:

[0153] ;

[0154] in, This indicates the physical distance between the leak point and the upstream data collection node. This represents the total length of the physical pipeline segment between the upstream and downstream data acquisition nodes. This indicates the propagation speed of pressure waves in the fluid medium of the pipe. Indicates time difference.

[0155] In the underlying implementation of the above parameter acquisition mechanism, the total length of the physical pipeline segment between the upstream and downstream acquisition nodes is... The analysis platform retrieves the sensor physical topology configuration table pre-stored in non-volatile memory. The propagation speed of the pressure wave in the pipe fluid medium is then calculated. It is physically related to fluid density, the elastic modulus of pipe material, and pipe wall thickness. The analysis platform reads this wave velocity value from a pre-set database of pipeline foundation physical parameters. This value is obtained through standard artificial flow calibration experiments by controlling valves during the commissioning phase after pipeline construction and before water filling and operation, and is stored as a constant.

[0156] Through the above calculations, the analysis platform achieves spatial positioning based on the propagation characteristics of hydrodynamics. Regarding the underlying algorithm implementation that uses Fast Fourier Transform to transform time-domain convolution into frequency-domain multiplication to reduce the computational complexity of the microprocessor during the cross-correlation function calculation, those skilled in the art can program and deploy it based on conventional digital signal processing theory. Its underlying time-frequency domain transformation and complex multiplication-addition operation mechanism are well-known technologies in this field and will not be elaborated upon here.

[0157] After obtaining the physical distance between the leak point and the upstream acquisition node, the analysis platform triggers a secondary calculation of the spatial coordinates based on a signal attenuation model. Due to the complex operating conditions of the fluid inside the pipeline, the fluid pressure wave is easily affected by local air pockets and pipe wall friction interference during propagation, leading to deviations in the location results from a single time-difference analysis. Introducing a location mechanism based on the attenuation characteristics of acoustic emission signals can create a synergistic and complementary multi-source physical quantity system with the fluid dynamics location mechanism.

[0158] When high-frequency acoustic emission waves propagate along the walls of a metal pipe, their wavefront energy decays exponentially with increasing propagation distance. The initial acoustic emission energy at the leak point propagates to both sides, and the amplitudes reaching the upstream and downstream sensing nodes depend on the unknown initial amplitude, the attenuation coefficient of the pipe wall medium, and the propagation distance, respectively. By establishing a ratio between the amplitudes of the signals received upstream and downstream, logarithmic operations can be used to eliminate the unknown initial amplitude variable in the equation, thus allowing the spatial distance parameter to be calculated in reverse.

[0159] The analysis platform retrieves acoustic emission feature vectors synchronously collected from the upstream and downstream edge gateways of the target pipe section. To eliminate single-point amplitude distortion caused by transient mechanical impact noise, the platform extracts the effective values ​​of the acoustic emission signal within a preset time window, using them as the signal amplitudes of the upstream and downstream nodes, respectively. The specific length of this preset time window is set to 0.1 seconds. Based on the exponential decay law of the acoustic emission signal in the medium, the platform constructs an algebraic equation. By performing logarithmic transformation and rearranging terms on the equation, the platform calculates the physical distance to the leakage point based on the attenuation model. The core formula for the signal attenuation model calculation is as follows:

[0160] ;

[0161] in, This represents the physical distance to the leak point calculated based on the attenuation model. This represents the total length of the physical pipeline segment between the upstream and downstream data acquisition nodes. Indicates the signal attenuation coefficient. The operator represents the natural logarithm operator. This represents the amplitude of the acoustic emission signal extracted by the upstream edge gateway. This indicates the amplitude of the acoustic emission signal extracted by the downstream edge gateway.

[0162] In the parameter fixing mechanism of the attenuation localization model, the signal attenuation coefficient The value of this coefficient is affected by the pipe material, pipe wall thickness, and the structure of the external anti-corrosion and insulation layer. During system initialization, the analysis platform reads this coefficient from a pre-set database of pipeline foundation physical parameters. The specific value of this coefficient is obtained during the pipeline construction phase through a standard lead-broken simulated acoustic emission experiment, and its range is set to 0.01–0.05 per meter.

[0163] After obtaining the physical distance between the leak point and the upstream acquisition node and the physical distance calculated based on the attenuation model, the analysis platform performs a multi-dimensional weighted fusion operation. The platform assigns corresponding weight coefficients to the two different location results. The weight allocation logic is set based on the sensor's hardware signal-to-noise ratio and historical fault statistical errors. Because the propagation stability of negative pressure waves is better than that of high-frequency acoustic emission signals, the platform sets the weight for time-difference location to 0.6 and the weight for attenuation location to 0.4. The platform then uses these weight coefficients to perform linear multiplication and addition operations to generate the final leak point coordinate output value. The core formula for comprehensive weighted location is as follows:

[0164] ;

[0165] in, This represents the overall weighted positioning distance. Indicates the time difference positioning weight. This indicates the physical distance between the leak point and the upstream data collection node. Indicates the attenuation of positioning weight. This represents the physical distance to the leak point calculated based on the attenuation model.

[0166] Through the aforementioned weighted collaborative calculation, the analysis platform eliminates random errors caused by single sensor data and outputs spatial coordinate positioning data. As for the underlying algorithms for logarithmic floating-point operations, register scheduling of multiply-accumulate instructions, and extraction of the effective value of the acoustic emission time-domain envelope within the microprocessor, those skilled in the art can develop them based on conventional digital signal processing principles and embedded system programming specifications. The underlying hardware operation logic is well-known technology in this field and will not be elaborated upon here.

[0167] After outputting the leakage status label and the comprehensive weighted location distance, the analysis platform triggers differentiated hierarchical alarms and scheduling strategies. To meet the timing requirements of industrial field control, a real-time operating system is deployed within the analysis platform to allocate different task priorities and computing resources based on the urgency of the leakage status label.

[0168] When the leakage status label corresponds to the normal operating status, the analysis platform uses background idle tasks to package and store the comprehensive fusion feature vector of the current data processing window and the status label into local non-volatile storage media for routine historical data archiving, maintaining the system's basic data recording cycle without additionally consuming network transmission bandwidth.

[0169] When the leakage status label corresponds to a micropore leakage, the analysis platform initiates a low-priority asynchronous network transmission task. The platform generates an early warning message, using a JSON data exchange format to serialize and encapsulate the current timestamp, sensor node physical address, leakage status label, and comprehensive weighted positioning distance. The platform then pushes the early warning message to the host computer system via industrial Ethernet using the application-layer MQTT protocol. Upon receiving the message, the host computer system parses the corresponding spatial coordinate data, generates a corresponding maintenance work order node in the pipeline network layer of the geographic information system, and assigns inspection personnel to perform leak detection and repair work.

[0170] When the leakage status label corresponds to a medium-level crack leakage, the analysis platform allocates a high-priority processing thread to execute the control logic. At the software communication level, the analysis platform sends an audible and visual alarm drive command containing precise spatial coordinates to the monitoring panel, alerting on-site operators. At the hardware bus level, the analysis platform utilizes the underlying controller area network bus or serial communication interface to issue register write commands to the upstream and downstream edge gateways of the target pipe section. This command directly overwrites the sampling timer configuration register of the analog-to-digital converter inside the edge gateway, setting the data sampling frequency to twice the original configuration parameter, increasing the number of physical parameter acquisition points per unit time, thereby obtaining high-resolution real-time fluctuation data to provide data support for potential pipe wall rupture trends.

[0171] When the leakage status label corresponds to a severe pipe wall rupture, there is a risk of a large-scale fluid leakage. The analysis platform employs a hardware interrupt response mechanism to execute interlocking control. To prevent transmission delays caused by conventional network protocol stacks under high data throughput, the analysis platform bypasses application layer network communication and directly pulls up the level of its onboard digital output pins to output a DC hardwired interlock signal. This interlock signal is directly connected to the digital input ports of the programmable logic controllers (PLCs) deployed at both ends of the pipe section. The PLCs respond to changes in the input port level, execute internally programmed emergency shutdown logic, and drive pneumatic actuators or electric valves to physically shut down the pipeline. The principle of physical shutdown is to prevent the fluid inside the pipe from continuing to leak due to its own weight and residual pressure at the pump station by cutting off the pipeline power source and the closed pipe section node, thereby controlling the scale of the accident.

[0172] After executing the physical shutdown command, the analysis platform starts a response confirmation timer, continuously monitoring the valve full-close limit feedback signal returned by the programmable logic controller (PLC) via the data bus. The analysis platform calculates the difference between the time of receiving the feedback signal and the hardware interrupt trigger time, using this as the response delay. If this response delay exceeds the set safety tolerance time limit, the analysis platform determines that there is mechanical jamming or electrical failure in the main control channel, and then switches to the backup relay channel to output a high-voltage shutdown command again, completing the closed loop of the fault handling logic. In the system parameter configuration, the analysis platform sets the specific value of this safety tolerance time limit to 3 seconds.

[0173] For mutex scheduling in real-time operating systems, byte alignment rules for message serialization encoding, and underlying ladder logic programming of programmable logic controllers, those skilled in the art can implement them according to conventional industrial automation control standards. The underlying software task scheduling and electrical control mechanisms are well-known technologies in this field and will not be elaborated here.

[0174] The physical environment of industrial pipeline networks is subject to electromagnetic interference and temperature fluctuations. Therefore, the analysis platform deploys a fault-tolerant redundancy mechanism to maintain the operation of the online monitoring system. This fault-tolerant redundancy encompasses hardware operation status monitoring, network communication link switching, and repair of missing sensing data.

[0175] At the hardware operation status monitoring level, the analysis platform incorporates a built-in hardware watchdog timer. During normal program operation, the main loop writes a reset instruction sequence to the watchdog timer's counter register at preset intervals, performing a watchdog feed operation. Electromagnetic interference in the environment may cause abnormal jumps in the microprocessor's program counter pointer, leading to an infinite loop in the software execution flow. If the main loop fails to complete the watchdog feed operation within the set overflow period, the hardware watchdog timer will pull low the microprocessor's hardware reset pin, causing the entire analysis platform to perform a cold start and restore the system to its initial calculation state. In the system configuration, the analysis platform sets the hardware watchdog timer's overflow period to 1.5 seconds.

[0176] At the network communication link switching level, the analysis platform is configured with dual physical layer communication interfaces: an industrial Ethernet interface as the primary link and a fourth-generation mobile communication technology (4G) wireless module as the backup link. The analysis platform periodically sends heartbeat status messages to the control center via the primary link. The transmission period for this heartbeat status message is set to 5 seconds. If no heartbeat confirmation frame is received from the control center for three consecutive cycles, the analysis platform determines that a physical disconnection has occurred in the wired network or a switch node failure has occurred. The analysis platform then cuts off the power supply to the peripheral devices of the Ethernet controller, activates the 4G wireless module, establishes a virtual private network tunnel, and redirects the transmission route of the leakage status tag and early warning message to the wireless backup link, ensuring the connectivity of the alarm channel under abnormal operating conditions.

[0177] In terms of missing data restoration, when a sensor on an edge gateway loses physical parameter data within a local time window due to loose wiring terminals, directly feeding these missing values ​​into a hybrid deep neural network will cause floating-point arithmetic anomalies or matrix dimension misalignment. Before performing batch normalization, the analysis platform uses cached historical time series data to perform first-order linear extrapolation to restore missing feature elements. The physical basis of first-order linear extrapolation is that pipeline fluid parameters exhibit continuous change inertia within short sampling periods, with their rate of change remaining stable. That is, between three adjacent discrete sampling points, assuming the physical quantity changes linearly with time, the difference between the current and previous times is equal to the difference between the previous two times. Based on this differential equality, a simplified equation is derived to predict the current value. The core formula for missing data restoration is as follows:

[0178] ;

[0179] in, This represents the feature element after repair at the current moment. Indicates the index of the current sampling time. This represents the historical feature elements of the same physical parameter collected at the previous sampling time. This indicates the index of the previous sampling time. This represents the historical feature elements of the same physical parameter collected at the first two sampling times. This indicates the index of the first two sampling times.

[0180] The analysis platform fills the corresponding index positions of the original data matrix with the repaired feature elements, reconstructs the complete feature vector sequence, and then inputs it into a one-dimensional convolutional neural network and a long short-term memory network to continue forward propagation inference.

[0181] The internal frequency divider logic of the hardware watchdog timer, the encrypted handshake protocol of the virtual private network tunnel, and the pointer recovery mechanism of the microprocessor memory stack can be developed by those skilled in the art according to conventional embedded system reliability design specifications. The underlying hardware fault tolerance and communication protocol stack management mechanism are well-known technologies in this field and will not be described in detail here.

[0182] Specific application examples:

[0183] Experimental verification and effect comparison:

[0184] To verify the effectiveness, accuracy, and reliability of the intelligent early leakage detection method for pipelines based on multi-parameter fusion proposed in this invention, a comprehensive simulation test was conducted on an industrial-grade annular pipeline experimental platform. The total length of the pipeline section on the experimental platform was set to... =500 meters, the pipe material is carbon steel. Monitoring nodes are set up along the line at 15-meter intervals to collect pressure, vibration, acoustic emission and temperature data simultaneously. Four working conditions were artificially set and simulated: "normal operation", "micro-pore leakage", "moderate crack leakage" and "severe pipe wall rupture", and a total of 8,000 sets of multi-dimensional data samples were collected as test benchmarks.

[0185] To verify the advantages of the hybrid deep neural network "one-dimensional convolutional neural network + long short-term memory network (1D-CNN + LSTM)" in the multi-source feature-level splicing and analysis platform executed by the edge gateway, a comparison experiment on the accuracy of leakage status identification was conducted. Different combinations of comparison models were set up, including:

[0186] The methods proposed in this invention include: single pressure sensor + support vector machine (SVM), single vibration sensor + conventional CNN, multi-source feature fusion + single LSTM, and multi-source feature fusion + hybrid deep neural network.

[0187] Experimental results (as attached) Figure 2 As shown in the figure:

[0188] When dealing with early and subtle features such as micropore leakage and moderate crack leakage, single sensors or traditional shallow machine learning models (such as SVM) are easily affected by environmental background noise, and the recognition accuracy hovers between 75% and 82%. While using "single vibration + CNN" or "multi-source fusion + LSTM" can improve the accuracy to about 85% to 91%, there is still a bottleneck in the recognition of micropores.

[0189] In contrast, the method of this invention extracts spatial local features through one-dimensional convolutional layers, extracts time-series features using long short-term memory networks, and, with the support of batch normalization processing, achieves an average recognition accuracy of over 98.5% for the four operating conditions. Especially in the most difficult-to-identify "micro-pore leakage state," the accuracy remains above 97.5%, demonstrating the powerful ability of multi-source physical parameter complementarity and hybrid deep neural networks to extract implicit features.

[0190] The model training convergence performance test verifies the effectiveness of the aforementioned cross-entropy loss function (Loss) and the introduction of the adaptive moment estimation optimization algorithm (Adam, parameter update formula) during closed-loop training operations performed on the analysis platform. The actual effect of the learning rate decay strategy.

[0191] The experiment recorded the trajectory of the model's loss value over 1000 iterations (epochs) (see attached). Figure 3(As shown). Experimental results show that for the traditional basic SGD optimization algorithm, due to the lack of momentum acceleration and adaptive learning rate adjustment, its loss value decreases slowly, and it gets stuck in local optima or oscillates in the later stages of iteration. In the end, the cross-entropy loss value stays at around 0.15, which cannot reach the set lower limit of error.

[0192] Conversely, in the early stages of training (0-200 epochs), the large initial global learning rate is beneficial. Guided by the first / second order momentum estimation and the value of 0.001, the cross-entropy loss of this invention rapidly drops from above 2.5 to below 0.5, demonstrating great corrective momentum.

[0193] In the later stages of training, as the adaptive decay strategy of the learning rate takes effect, the loss curve effectively avoids oscillations in the minimum region. At approximately 450 epochs, the cross-entropy loss value of the model in this invention converges smoothly and penetrates the set lower error limit of 0.01, eventually stabilizing at around 0.004, demonstrating the efficiency and stability of the underlying gradient descent and parameter update mechanism adopted in this invention.

[0194] Comparison of positioning accuracy using multi-source information weighting:

[0195] To verify the accuracy of the aforementioned multi-source information weighted positioning algorithm based on time difference analysis and signal attenuation model, leakage was triggered on pipe sections within a range of 50 to 450 meters from the upstream sensor node (with a test node set at 50-meter intervals), and the absolute distance error (the absolute value of the difference between the actual distance and the calculated distance) calculated by different algorithms was recorded.

[0196] The comparison algorithms include:

[0197] Distance is calculated solely based on the time difference of the negative pressure wave (corresponding to the aforementioned formula). ), and calculate the distance solely based on the acoustic emission attenuation model (corresponding to the aforementioned formula). ), and the comprehensive weighted positioning calculation proposed in this invention (corresponding to the core formula) The weights are 0.6 and 0.4 respectively.

[0198] Experimental test results (as attached) Figure 4 As shown in the figure:

[0199] The error of single negative pressure wave time difference positioning gradually increases when transmitted over long distances due to airbag obstruction (maximum error of about 8.5 meters).

[0200] Single attenuation positioning is easily affected by sudden changes in damping of the local anti-corrosion layer, leading to measurement distortion (the overall error fluctuates between 4 and 7 meters).

[0201] The heterogeneous weighted fusion mechanism adopted in this invention neutralizes the high-frequency random errors of fluid dynamics and acoustics, and the absolute positioning error at each test node is suppressed to within 1.5 meters, which meets the precise positioning requirements for pipeline leakage excavation and repair in industrial sites and improves the system's spatial coordinate secondary solution capability.

Claims

1. A pipeline early leakage detection system based on a multi-sensor network, characterized in that, The system includes: A multi-sensor network deployed on a pipeline, the multi-sensor network comprising multiple monitoring nodes deployed along the pipeline to be inspected, each monitoring node being fixedly equipped with a pressure sensor, a vibration sensor, an acoustic emission sensor and a temperature sensor, for acquiring the corresponding pressure signal, vibration signal, acoustic emission signal and temperature signal; An edge gateway, electrically connected to the multi-sensor network, is used to receive the pressure signal, vibration signal, acoustic emission signal, and temperature signal and perform preprocessing to generate standardized preprocessed signals corresponding to each physical dimension. The first processor, connected to the edge gateway, is used to extract features from the standardized preprocessed signal based on a sliding time window, and to construct and generate pressure feature vector, vibration feature vector, acoustic emission feature vector and temperature feature vector. The second processor, electrically connected to the first processor, is used to perform spatiotemporal alignment and splicing of the pressure feature vector, vibration feature vector, acoustic emission feature vector and temperature feature vector to generate a comprehensive fusion feature vector and transmit it to the analysis platform. The analysis platform is used to input the integrated feature vector into a hybrid deep neural network for inference and judgment, and output the leakage status label of the pipeline to be detected; When the leakage status label indicates that a leakage exists, the analysis platform triggers the leakage location program and executes alarm and scheduling strategies based on the leakage status label.

2. The system according to claim 1, characterized in that, The edge gateway includes an adaptive filter, and the edge gateway performs preprocessing, including: The time-domain signal sequences of the pressure signal, vibration signal, acoustic emission signal and temperature signal are extracted and subjected to discrete wavelet decomposition to obtain low-frequency approximation coefficient components and high-frequency detail coefficient components. An adaptive soft thresholding algorithm is used to perform threshold quantization on the high-frequency detail coefficient components. The retained low-frequency approximation coefficient components and the quantized high-frequency detail coefficient components are used to reconstruct a one-dimensional preliminary denoised signal in the time domain. The background signal collected by the environmental reference sensor deployed in the non-monitoring area is used as the input reference noise signal vector, and is input together with the preliminary denoising signal into the minimum mean square adaptive filter; By iteratively optimizing and dynamically adjusting the tap weight coefficients, the statistical characteristics of the residual noise are approximated, and the residual noise is separated from the initial denoised signal, outputting the physical signal after composite denoising. The normalized preprocessed signal is generated by performing maximum and minimum value normalization processing on the physical signal.

3. The system according to claim 1, characterized in that, The first processor is specifically used for: The mean, variance, peak-to-peak value, and kurtosis parameters are calculated and extracted from the standardized preprocessed signal corresponding to the pressure signal, and then combined according to the set arrangement order to construct a one-dimensional pressure feature vector. Fast Fourier Transform is performed on the standardized preprocessed signal of the corresponding vibration signal to extract the centroid frequency. Wavelet packet transform is used to calculate the wavelet packet energy entropy. The centroid frequency and wavelet packet energy entropy are concatenated and combined in a set order to construct a one-dimensional vibration feature vector. The standardized preprocessed signal of the corresponding acoustic emission signal is set with a dynamic trigger threshold calculated based on the mean of the signal envelope curve to identify acoustic emission transient events. Impact count parameters, average event energy parameters and average peak amplitude parameters are extracted and combined in a set order to construct a one-dimensional acoustic emission feature vector. The basic temperature values ​​of each distributed temperature sensor node are calculated using the standardized preprocessed signal of the corresponding temperature signal. The spatial maximum temperature difference parameter, spatial temperature gradient parameter, and spatial coefficient of variation parameter are extracted and combined according to the set arrangement order to construct a one-dimensional temperature feature vector.

4. The system according to claim 3, characterized in that, The identification of the acoustic emission transient event includes: The envelope curve is obtained by performing envelope detection on the standardized preprocessed signal of the corresponding acoustic emission signal using Hilbert transform; The dynamic trigger threshold is obtained by multiplying the arithmetic mean of the values ​​of all sampling points of the envelope curve in the current data processing window with the set trigger coefficient. When the value of the envelope curve increases from small to large and crosses the dynamic trigger threshold, the sampling number corresponding to the crossing point is marked as the event start point. When the value of the envelope curve decreases from large to small and crosses the dynamic trigger threshold within the set event lock time window and does not rise again by crossing the dynamic trigger threshold, the sampling number corresponding to the falling point is marked as the event end point. The envelope curve data segment between the event start point and the event end point constitutes an independent acoustic emission transient event.

5. The system according to claim 1, characterized in that, The second processor is specifically used for: The global timestamp label is obtained from the edge gateway. The global timestamp label is generated by the edge gateway when reading each feature vector to generate the standardized preprocessed signal. The pressure feature vector, vibration feature vector, acoustic emission feature vector, and temperature feature vector are aligned in the time dimension using the system clock cycle of the current data processing window as a reference criterion. The spatial sequence index is obtained from the edge gateway, and the spatial sequence index is generated by the edge gateway according to the distribution of the acquisition nodes along the axis of the pipeline to be detected. The pressure feature vector, vibration feature vector, and acoustic emission feature vector belonging to the same monitoring section coordinate are clustered into a local physical parameter set. The temperature feature vector reflecting the heat conduction state of the entire pipeline is added as a global physical parameter to the back region of the local physical parameter set. The local physical parameter set and the global physical parameter are merged by performing a cascade mapping operation through the data bus to generate a cross-dimensional high-dimensional data array as the comprehensive fusion feature vector. The integrated feature vector is sent to the analysis platform, which is further used for: The global moving average is read from the parameter weight file of the deep neural network model as the batch mean, and the global moving variance is read as the batch variance. The batch mean and batch variance are used to standardize the integrated feature vector, and scaling and translation parameters are introduced to perform linear reconstruction, outputting the batch normalized feature vector.

6. The system according to claim 5, characterized in that, The hybrid deep neural network includes a one-dimensional convolutional neural network, a long short-term memory network, and a fully connected layer. The analysis platform is specifically used for: The one-dimensional convolutional neural network performs sliding matrix multiplication and addition operations on the batch-normalized feature vectors according to the convolutional kernel of a set length to generate local feature mapping values. Max pooling dimensionality reduction is then performed on the output local feature mapping matrix to generate a feature sequence. The feature sequence is received through the Long Short-Term Memory network, the spatial structure dependency is resolved through the internal control mechanism, the hidden layer output vector of the last sequence step is obtained and flattened and passed to the fully connected layer; The high-dimensional abstract features are mapped to classification nodes equal to the set number of leakage categories through the fully connected layer. The output values ​​of the classification nodes are probabilistically mapped by the normalized exponential function, and the predicted probability of the pipeline to be detected being in each leakage state is calculated to obtain a set of probability distributions. Extract the state label corresponding to the maximum predicted probability in the probability distribution set as the leakage state label of the pipeline to be detected.

7. The system according to claim 6, characterized in that, Before extracting the state label corresponding to the maximum predicted probability, the analysis platform is also used for: The maximum predicted probability value in the probability distribution set is compared with the set confidence threshold. If the maximum predicted probability value is greater than or equal to the confidence threshold, then the corresponding category index is output as the leakage status label of the pipeline to be detected. If the maximum predicted probability value is less than the confidence threshold, the output result is marked as undecided, the moving average determination mechanism is started, the probability distribution set of the next two consecutive adjacent data processing windows is obtained, and the probability distribution set of the current data processing window is combined with the probability distribution set of the current data processing window to form a moving determination sequence with a total length of 3 windows. The arithmetic mean of the predicted probability of the same category is calculated to generate a smoothed probability distribution set. The maximum smoothed predicted probability value in the smoothed probability distribution set is compared with the set secondary decision threshold. If the maximum smoothed predicted probability value is greater than or equal to the secondary decision threshold, its corresponding status label is output. If the value is still less than the secondary judgment threshold, the leakage status label of the pipeline to be tested will be forcibly output as normal operation status.

8. The system according to claim 1, characterized in that, The analysis platform triggers a leak location procedure, including: Retrieve upstream and downstream pressure feature vectors and upstream and downstream acoustic emission feature vectors synchronously collected by edge gateways located upstream and downstream of the target pipe segment to be detected; Time alignment analysis is performed on the upstream and downstream pressure feature vectors to obtain the time difference. The time difference between the leak point and the upstream acquisition node is calculated by combining the propagation speed of the pressure wave in the pipeline fluid medium to locate the physical distance. Based on the upstream and downstream acoustic emission feature vectors, the effective values ​​of the acoustic emission signals within the set time window are extracted and used as the signal amplitudes of the upstream and downstream nodes, respectively. Utilizing the physical mechanism that the energy of the acoustic emission wavefront attenuates with the increase of propagation distance, an algebraic equation is constructed and solved using the signal amplitudes of the upstream and downstream nodes and the acoustic emission attenuation coefficient as parameters, and the attenuation positioning physical distance calculated based on the attenuation model is obtained. Time difference positioning weight and attenuation positioning weight are assigned to the physical distance of time difference positioning and the physical distance of attenuation positioning, respectively. A linear multiplication-addition operation is performed to generate a comprehensive weighted positioning distance, which is used as the final output value of the leak point coordinates.

9. The system according to claim 8, characterized in that, The analysis platform is connected to the host computer control center. The analysis platform executes alarm and scheduling strategies based on the leakage status label, including: When the leakage status label corresponds to the micropore leakage status, a first-level early warning message is generated and sent to the host computer control center. When the leakage status label corresponds to the medium crack leakage status, an audible and visual alarm drive command is sent to the host computer control center, and commands are sent to the upstream and downstream edge gateways to set the data sampling frequency to twice the original setting parameter. When the leakage status label corresponds to a severe pipe wall rupture, a three-level emergency interlock mechanism is triggered, outputting a DC hard-wired interlock signal. Based on the interlock signal, the pneumatic actuator or electric valve is driven to perform a physical shut-off operation to isolate the pipeline, and a response confirmation timer is started to continuously monitor the valve full-close limit feedback signal, thus completing the closed loop of the fault handling logic.

10. The system according to claim 5, characterized in that, The analysis platform is also used for: Before performing the batch normalization process, when the physical parameter acquisition value is lost within a local time window, historical feature elements of the same physical parameter acquired at the previous sampling time and the previous two sampling times are extracted. Establish the differential equality relationship between three adjacent discrete sampling points, generate the prediction equation for the current time value, and obtain the repaired feature elements at the current time through first-order linear extrapolation repair; The repaired feature elements at the current moment are filled into the corresponding index positions of the original data matrix to reconstruct the complete feature vector sequence; The complete feature vector sequence is substituted into the integrated feature vector to continue the batch normalization process.

Images